A Frequency-Domain Neural-Network Model for High-Power RF Transistors

Abstract

In power amplifier design, when equivalent-circuit models are not available, behavioral models present a possible solution to represent the nonlinear behavior of the transistor. Among the existing behavioral models, the interpolation capabilities of the artificial neural networks have been explored to successfully approximate the measured load-pull behavior of such devices. However, these models require a large set of measurements of the device, that, in practice, are not always available. Typically, these models rely on power swept load-pull measurements and, since the PA design is nowadays targeting several carrier frequencies, the number of power levels and loads cannot be very large. This normally leads to unreasonable results when the model is implemented in a circuit simulator, especially at small power levels. This article proposes a simple solution to that problem, by taking an artificial neural network-based model and creating virtual, low power, load-pull data from the S-parameters of the device.